JP3137550B2 - Audio encoding / decoding device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method of determining pitch frequency, voiced / unvoiced judgment of each harmonic, and spectrum amplitude information from speech.
A speech encoding / decoding device that performs encoding / decoding using two parameters, wherein the spectrum amplitude information is represented by L
The present invention relates to an audio encoding / decoding device that is expressed by PC (linear prediction) coefficients and encodes / decodes its parameters efficiently.

[0002]

2. Description of the Related Art In recent years, with the development of digital signal processing technology, digital communication services have been diversified, and there has been an increasing demand for lower bit rates due to limitations on transmission capacity in communication. High-efficiency speech coding technology is an indispensable technology to meet the demand. Pitch frequency, voiced / unvoiced judgment of each harmonic, spectrum amplitude information
MBE (Multi Band Excit) that encodes two parameters
ed) The coding method is known as an excellent coding method capable of obtaining good sound quality even at a low bit rate (IEEE
Trans ASSP VOL 36. NO.8. 1988). Also, an MB using an LPC coefficient as a parameter representing spectrum amplitude information
An E encoding method is also known (Electronics Letters Vol.2
7. No.14. 1991).

[0003] An encoding / decoding apparatus using an MBE encoding method using LPC coefficients as parameters representing spectral amplitude information will be described below. First, a conventional encoding device will be described with reference to FIG. In FIG. 8, reference numeral 1 denotes a pitch frequency estimating unit that receives an input voice signal and outputs a pitch frequency. Reference numeral 2 denotes a voiced / unvoiced determination unit which receives an input voice signal and a pitch frequency and outputs voiced / unvoiced determination of each harmonic. Reference numeral 3 denotes an LPC analysis unit that receives an input audio signal and outputs an LPC coefficient. Reference numeral 4 denotes a quantizer that receives LPC coefficients as input and outputs quantized LPC coefficients as output. Reference numeral 5 denotes an LPC inverse filter that receives an input speech signal and LPC coefficients as inputs and outputs an LPC prediction residual. Reference numeral 6 denotes a frame power calculation unit that receives the LPC prediction residual and inputs the frame power. 7
Is a quantizer that receives frame power as input and outputs quantized frame power. A multiplexer 8 receives a pitch frequency, a voiced / unvoiced determination, a quantized LPC coefficient, and a quantized frame power as inputs and outputs a code.

Next, the operation will be described with reference to FIG. The pitch frequency estimating unit 1 calculates a pitch frequency from an input voice signal. As a means for calculating the pitch frequency, a method using a correlation function or a spectrum amplitude of an input voice signal has been conventionally known. Next, the voiced / unvoiced determination unit 2 calculates the spectrum of the input voice signal,
Harmonic frequencies are determined based on the pitch frequency, and voiced / unvoiced determination of each harmonic is performed based on the harmonic frequencies. As a voiced / unvoiced determination method of each harmonic, a method of performing determination based on a difference between a spectrum when each harmonic is assumed to be voiced and a spectrum of an input voice signal is conventionally known. The LPC analysis unit 3 performs an LPC analysis of the input audio signal,
Calculate the coefficient. In the quantizer 4, the LPC coefficients are quantized by an efficient quantization method such as conventionally known vector quantization. The LPC inverse filter 5 passes the input audio signal through the LPC inverse filter using the LPC coefficients. As a result, an LPC prediction residual of the input audio signal is obtained. The power calculator 6 calculates an average value of the power of the LPC prediction residual within the frame, and outputs the average value as the frame power. In the quantizer 7, the frame power is quantized by an efficient quantization method as conventionally known. The multiplexer 8 efficiently encodes the pitch frequency, the voiced / unvoiced determination of each harmonic, the quantized LPC coefficient, and the quantized frame power obtained as information to be sent to the decoding device. As a result, the encoded pitch frequency, voiced / unvoiced decision, LPC coefficient, and frame power are obtained as outputs of the encoding device.

Next, a conventional decoding apparatus will be described with reference to FIG. In FIG. 9, reference numeral 11 denotes a demultiplexer which receives a code and outputs pitch frequency, voiced / unvoiced determination, quantized frame power, and quantized LPC coefficient. 12, a pitch frequency and a voiced / unvoiced judgment are input,
A mode determination unit that outputs the MBE combining mode. 13
Is the input of the quantized LPC coefficient and the inversely quantized LPC
This is an inverse quantizer that outputs coefficients. Reference numeral 14 denotes an input of the inversely quantized LPC coefficient, and a spectrum-emphasized LP.
This is a spectrum emphasis unit that outputs the C coefficient. Reference numeral 15 denotes a harmonic amplitude multiplier calculation unit that receives the spectrum-emphasized LPC coefficient and outputs the amplitude multiplier of each harmonic. 1
Reference numeral 6 denotes an inverse quantizer which receives the quantized frame power as input and outputs the inversely quantized frame power as output. Reference numeral 17 denotes a sound source generation unit which receives a pitch frequency, voiced / unvoiced judgment, and frame power as input and outputs a sound source signal. 18
Is an MBE synthesizing unit which receives an MBE synthesis mode, a sound source signal, and each harmonic amplitude multiplier as inputs, and outputs decoded speech.

Next, the operation will be described with reference to FIG. First, the demultiplexer 11 extracts a pitch frequency, voiced / unvoiced determination, quantized frame power, and quantized LPC coefficients from the code sent from the encoding device side. The mode determination unit 12 determines the MBE combining mode by a conventionally used mode determination method based on the pitch frequency and the voiced / unvoiced determination. The inverse quantizer 13 inversely quantizes the quantized LPC coefficients. The spectrum emphasizing unit 14 performs conventionally known processing such as formant emphasis and high-frequency emphasis on the LPC coefficients in order to improve the subjective sound quality of the output sound. The harmonic amplitude multiplier calculator 15 converts the LPC coefficient into a frequency domain by FFT. The spectrum envelope of the synthesized speech signal can be obtained by obtaining the absolute value of each of the obtained Fourier coefficients and taking the reciprocal. The amplitude multiplier of each harmonic of the synthesized voice signal is a frequency n which is an integral multiple of the pitch frequency w0.
It is obtained by finding w0 and extracting its frequency components from the spectral envelope. The inverse quantizer 16 inversely quantizes the quantized frame power. Sound source generator 17
Then, a sound source is generated based on the pitch frequency, voiced / unvoiced judgment, and frame power. To generate a sound source, first, the frequency of each harmonic is obtained from the pitch frequency, and the frequency of each harmonic is calculated. In addition, a basic sound source signal is generated for each harmonic frequency based on the voiced / unvoiced determination. Conventionally, a method is known in which a sine wave is used if the harmonic is voiced, and a white noise is used if the harmonic is unvoiced. Next, from frame power,
A sound source signal is generated by estimating a sound source amplitude of each harmonic and multiplying the estimated sound source amplitude by a basic sound source signal. MBE synthesis unit 18
Then, based on the sound source signal generated in this way, the amplitude of each harmonic, and the MBE synthesis mode, conventionally known MBE synthesis is performed, and a decoded audio signal is output.

[0007]

However, in the above-mentioned conventional speech coding / decoding apparatus, since the power of each harmonic is calculated from the spectral envelope obtained only from the LPC coefficient, it cannot be represented by the LPC coefficient. Since such a finer spectral envelope cannot be obtained, there is a problem that the power of each harmonic is not accurately encoded, and the quality of the decoded speech is reduced. In addition, even if the power of the input audio signal greatly changes within a frame, there is no means to represent the power, and the input audio signal cannot follow the input signal. I was

[0008] The present invention solves the above-mentioned conventional problems, and provides an excellent speech encoding and decoding method for realizing high decoded speech quality.
It is an object to provide a decoding device.

[0009]

According to the present invention, in order to achieve the above object, a coding apparatus includes a pitch frequency estimating unit,
A voiced / unvoiced determination unit for voiced / unvoiced determination of each harmonic, and LP
C (linear prediction) analysis unit and a quantum for quantizing LPC coefficients
, LPC inverse filter, and LPC prediction residual
At least the residual information other than the residual frame power
Difference information calculator, pitch frequency, voiced / unvoiced judgment and quantization
To encode the obtained LPC coefficients and residual information
And the decoding device, from the code to the pitch frequency and
The voiced / unvoiced decision, the quantized LPC coefficient and the residual information
Determine the demultiplexer to decode and the MBE combining mode
Mode determining unit to determine the LPC coefficient quantized
An inverse quantizer for extracting LPC coefficients by converting
Harmonic amplitude multiplier to find the amplitude multiplier for each harmonic from the number
Output part, pitch frequency, voiced unvoiced frequency, residual information, etc.
A sound source generation unit for generating a sound source signal from the
Wherein the residual information calculation unit in the encoding device is
Frame power calculation unit and a quantum that quantizes the frame power
And a prediction residual normalization unit for normalizing the LPC prediction residual
And the power spectrum for calculating the power spectrum of the prediction residual
Calculate the power of harmonics contained in the residual by the vector calculator
From the residual harmonic power calculator and the residual harmonic power
Spectral correction vector for calculating the spectral correction vector
Torque calculator and the amount to quantize the spectral correction vector
And a sound generator in the decoding apparatus.
The frame power and spec
A sound source is generated using the vector
The decoding frequency and the decoding frequency are respectively
The LPC coefficient and the spectrum correction
Calculate bit allocation for each frame
A quantization level number calculation unit is provided, and LP
Quantizer / dequantizer for quantizing / dequantizing C coefficient
Amount to quantize and dequantize spectral correction vector and spectrum correction vector
The quantizer and inverse quantizer are each a quantization level calculation unit.
At different quantization levels for each frame given by
Multiplexing and inverse quantization, and multiplexer and demultiplexer
Each of the multiplexors is calculated by the quantization level calculation unit.
LPC coefficient and quantity based on the assigned bit allocation
Encode and decode child spectral correction vectors
It was made .

[0010]

According to the above arrangement, the encoding apparatus uses the LPC
By extracting the power of each harmonic from the prediction residual and generating a sound source signal based on the power of each harmonic in the decoding device, the power of each harmonic can be accurately represented.

[0011] Further, the encoding device extracts a power change in the frame from the LPC prediction residual, and the decoding device generates an excitation signal based on the power change in the frame. It can follow changes in power within the frame.

[0012]

【Example】

Embodiment 1 Hereinafter, a first embodiment of the present invention will be described with reference to the drawings. First, an encoding device will be described with reference to FIG. In FIG. 1, reference numeral 101 denotes a pitch frequency estimating unit that receives an input speech signal and outputs a pitch frequency. Reference numeral 102 denotes a voiced / unvoiced determination unit which receives an input voice signal and a pitch frequency and outputs voiced / unvoiced determination of each harmonic. An LPC analysis unit 103 receives an input audio signal and outputs LPC coefficients. 10
Reference numeral 4 denotes a quantizer that receives LPC coefficients as input and outputs quantized LPC coefficients as output. 105 is the input audio signal and L
LP with PC coefficient as input and LPC prediction residual as output
C is an inverse filter. Reference numeral 106 denotes a frame power calculation unit that receives the LPC prediction residual and outputs the frame power. A quantizer 107 receives the frame power as input and outputs the quantized frame power as output. 108 is L
A prediction residual normalization unit that receives a PC prediction residual and a frame power as inputs and outputs a normalized prediction residual as an output. Reference numeral 109 denotes a residual power spectrum calculator that receives the normalized prediction residual as an input and outputs the power spectrum of the prediction residual as an output. 1
Reference numeral 10 denotes a residual harmonic power calculator that receives the power spectrum of the prediction residual and the pitch frequency as inputs and outputs the power of each harmonic of the residual. Reference numeral 111 denotes a spectrum correction vector calculation unit which receives the power of each harmonic of the residual and outputs a spectrum correction vector. A quantizer 112 receives a spectrum correction vector as an input and outputs a quantized spectrum correction vector as an output. Reference numeral 113 denotes a residual information calculation unit including elements from the frame power calculation unit 106 to the quantizer 112. 114 is a calculated pitch frequency, voiced / unvoiced determination, quantized LPC coefficient,
This is a multiplexer that receives a quantized frame power and a quantized spectrum correction vector and outputs a code.

Next, the operation will be described with reference to FIG. The pitch frequency estimating unit 101 calculates the pitch frequency from the input voice signal. As a means for calculating the pitch frequency, a method using a correlation function or a spectrum amplitude of an input voice signal has been conventionally known. Next, the voiced / unvoiced determination unit 102 calculates the spectrum of the input voice signal, obtains the harmonic frequencies based on the pitch frequency, and performs voiced / unvoiced determination of each harmonic based on these. As a voiced / unvoiced determination method of each harmonic, a method of performing determination based on a difference between a spectrum when each harmonic is assumed to be voiced and a spectrum of an input voice signal is conventionally known. L
The PC analysis unit 103 performs an LPC analysis on the input audio signal and calculates an LPC coefficient. In the quantizer 104, LP
Quantize the C coefficient. As a method of quantizing LPC coefficients,
2. Description of the Related Art A method of converting an LPC coefficient into an LSP coefficient and quantizing the LSP coefficient by an efficient quantization method such as vector quantization is conventionally known. In the LPC inverse filter 105,
The input audio signal is passed through an LPC inverse filter using the LPC coefficients. As a result, an LPC prediction residual of the input audio signal is obtained.

The power calculation section 106 calculates an average value of the power of the LPC prediction residual within the frame and outputs the average value as the frame power. In the quantizer 107, the frame power is quantized by an efficient quantization method as conventionally known. The prediction residual normalization unit 108 normalizes the LPC prediction residual using the frame power. The residual power spectrum calculator 109 performs fast Fourier transform on the normalized LPC prediction residual. At this time, it is desirable to multiply the LPC prediction residual by a window function such as a Hamming window in order to reduce the influence on the spectrum due to a sudden change in the signal waveform at both ends of the frame. The power spectrum of the prediction residual is obtained by squaring each of the obtained Fourier coefficients. The residual harmonic power spectrum calculator 110 calculates the power of the L harmonics included in the LPC prediction residual. The power of harmonics can be calculated by finding L frequencies nw0 (1 ≦ n ≦ L) that are integer multiples of the pitch frequency w0 and extracting the frequency components from the power spectrum of the prediction residual. In addition, the entire band wc of the audio signal is divided into L number of harmonics, and the average value of the power spectrum of the prediction residual is calculated in each band. Even if an error occurs between the frequency of the wave and a frequency that is an integral multiple of the obtained pitch frequency, the power of each harmonic can be more accurately estimated. The spectrum correction vector calculation unit 111 calculates a spectrum correction vector from the power of each harmonic of the obtained prediction residual. Harmonic number L
Changes depending on the pitch frequency w0. In order to quantize this efficiently, for example, out of L harmonics, a harmonic in an acoustically important band is fixed regardless of the number L of harmonics. The number N may be selected and used as a spectrum correction vector, and the quantized N spectral correction vectors may be quantized by the quantizer 112 using N-dimensional vector quantization. The multiplexer 114 efficiently encodes the pitch frequency, the voiced / unvoiced determination of each harmonic, the quantized LPC coefficient, the quantized frame power, and the quantized spectrum correction information obtained as information to be sent to the decoding device.
As a result, the encoded pitch frequency, voiced / unvoiced decision, LPC coefficient, frame power, and spectrum correction information are obtained as the output of the encoding device.

Next, the decoding apparatus will be described with reference to FIG. In FIG. 2, reference numeral 201 denotes a demultiplexer that receives a code and outputs a pitch frequency, a voiced / unvoiced determination, a quantized frame power, a quantized spectrum correction vector, and a quantized LPC coefficient. 202 is a pitch frequency,
A mode determination unit that receives a voiced / unvoiced determination and outputs an MBE combining mode. An inverse quantizer 203 receives the quantized LPC coefficients as input and outputs the inversely quantized LPC coefficients as output. Reference numeral 204 denotes a spectrum emphasizing unit that receives the inversely quantized LPC coefficients and outputs the spectrum-emphasized LPC coefficients. 205 is a spectrum emphasized L
It is a harmonic amplitude multiplier calculation unit that receives a PC coefficient as input and outputs an amplitude multiplier of each harmonic. An inverse quantizer 206 receives the quantized frame power as input and outputs the inversely quantized frame power as output. An inverse quantizer 207 receives a quantized spectrum correction vector as an input, and outputs an inversely quantized spectrum correction vector as an output. Reference numeral 208 denotes a sound source generation unit that receives a pitch frequency, a voiced / unvoiced determination, a frame power, and a spectrum correction vector, and outputs a sound source signal. Reference numeral 209 denotes an MBE synthesizing unit that receives an MBE synthesis mode, a sound source signal, and each harmonic amplitude multiplier as inputs and outputs a decoded speech.

Next, the operation will be described with reference to FIG. First, in the demultiplexer 201, the pitch frequency, voiced / unvoiced judgment,
The quantized frame power, the quantized spectrum correction vector, and the quantized LPC coefficient are respectively extracted. The mode determination unit 202 determines the MBE combining mode by a conventionally used mode determination method based on the pitch frequency and the voiced / unvoiced determination. The inverse quantizer 203
Dequantize the quantized LPC coefficient. The spectrum emphasizing unit 204 performs conventionally known processing such as formant emphasis and high-frequency emphasis on the LPC coefficients in order to improve the subjective sound quality of the output sound. When there is no need to perform such spectrum enhancement, there is no need to provide this spectrum enhancement unit. The harmonic amplitude multiplication calculation unit 205 converts the LPC coefficient into a frequency domain by FFT. At this time, it is desirable to multiply by a window function such as a Hamming window in order to reduce the influence on the spectrum due to a sudden change in the signal waveform at both ends of the frame.
The spectrum envelope of the synthesized speech signal can be obtained by obtaining the absolute value of each of the obtained Fourier coefficients and taking the reciprocal. The amplitude multiplier of each harmonic of the synthesized speech signal is obtained by obtaining a frequency nw0 that is an integral multiple of the pitch frequency w0 and extracting the frequency component from the spectrum envelope. Further, if the entire band wc of the audio signal is divided into L number of harmonic bands, and the average value of the spectrum envelope is calculated in each band, and the average value is used as the amplitude multiplier of each harmonic, the frequency of the harmonic of the audio is calculated. Even if an error occurs between the calculated pitch frequency and an integral multiple of the pitch frequency, the amplitude multiplier of the harmonic can be obtained more accurately. The inverse quantizer 206 inversely quantizes the quantized frame power. The inverse quantizer 207 inversely quantizes the quantized spectrum correction vector.

In the sound source generator 208, the pitch frequency and
A sound source is generated based on the voiced / unvoiced determination, the frame power, and the spectrum correction vector. To generate a sound source, first, the frequency of each harmonic is obtained from the pitch frequency, and the frequency of each harmonic is calculated. In addition, a basic sound source signal is generated for each harmonic frequency based on the voiced / unvoiced determination. Conventionally, a method is known in which a sine wave is used if the harmonic is voiced, and a white noise is used if the harmonic is unvoiced. Next, the sound source amplitude of each harmonic is estimated from the frame power and the spectrum correction vector, and the sound source signal is generated by multiplying the estimated sound source amplitude by the basic sound source signal. As means for calculating the excitation amplitude of each harmonic, first, the coding apparatus calculates the spectrum power of each harmonic included in the residual from the spectrum correction vector and multiplies it by the frame power. Theoretically, the spectrum of the audio signal is reproduced by this result and the spectral envelope. However, when performing MBE synthesis, the modeling method differs between voiced and unvoiced for each harmonic as described above. , And voiced / unvoiced judgment models, it is desirable to remove the effect of the window function when calculating the spectrum. As a method of removing the influence, a method of deconvolving the spectrum of the window function by assuming respective spectrum models for the amplitude of the sound source obtained from the frame power and the spectrum correction vector can be considered. The MBE combining unit 209 performs conventionally known MBE combining based on the thus generated sound source signal, the amplitude of each harmonic, and the MBE combining mode, and outputs a decoded audio signal.

As described above, according to the present embodiment, the conventional encoding apparatus includes the prediction residual normalizing unit 108, the residual power spectrum calculating unit 109, and the residual harmonic power calculating unit 1
10, a spectrum correction vector calculation unit 111, and a quantizer 112 for quantizing the spectrum correction vector to calculate and code a parameter representing a fine spectrum of each harmonic that could not be removed by the LPC analysis. And can be transmitted as information to the decoding device. Therefore, by providing the excitation device 208 in the decoding device in consideration of the spectrum correction vector, the spectrum of the harmonic included in the LPC prediction residual can be corrected at the time of decoding. The spectrum of the wave can be represented more accurately, and the sound quality can be improved.

(Embodiment 2) Next, a second embodiment of the present invention will be described with reference to FIG. FIG. 3 shows the configuration of the encoding apparatus according to the present embodiment. Elements common to those of the encoding apparatus according to the first embodiment shown in FIG. 1 are denoted by the same reference numerals, and redundant description is omitted. In this embodiment,
The decoding device has the same configuration as the decoding device of the first embodiment shown in FIG.

In FIG. 3, reference numeral 115 denotes an inverse quantizer which receives quantized LPC coefficients as input and outputs inversely quantized LPC coefficients as output. An inverse quantizer 116 receives the quantized frame power as input and outputs the inversely quantized frame power as output. Residual information calculation unit 113A in the present embodiment
Indicates that the residual information calculation unit 113 shown in FIG.
16 is added. The inverse quantizer 115 inversely quantizes the LPC coefficients quantized by the quantizer 104. As a result, the LPC coefficient obtained as the output of the inverse quantizer has a value exactly equal to the output of the inverse quantizer 207 in the decoding device of FIG. The error caused by this quantization is reflected on the prediction residual output from the inverse filter 105. Similarly, the inverse quantization 116 inversely quantizes the frame power quantized by the quantizer 107. The inversely quantized frame power has a value equal to the output of the inverse quantizer 206 in the decoding device of FIG. The error generated by the quantization is reflected in the normalized prediction difference output from the prediction residual normalization unit 108. Therefore, these quantization errors are calculated by the residual power spectrum calculator 109 and the residual harmonic power calculator 11.
0, which is calculated by the spectrum correction vector calculation unit 111 and reflected in the output spectrum correction vector. That is, the number of coded bits is reduced and the quantizer 104
Alternatively, even if the quantization error of the quantizer 107 or both becomes large, the error can be absorbed by appropriately assigning bits to the quantizer 112.

As described above, according to the present embodiment, the encoding device includes the inverse quantizer 1 for inversely quantizing the quantized LPC coefficient.
15 and an inverse quantizer 116 for inversely quantizing the quantized frame power.
By operating the PC inverse filter 105 and operating the prediction residual normalization unit 108 with the inversely quantized frame power, those quantization errors can be corrected by the spectrum correction vector, and the LPC coefficient or the frame power Alternatively, even if the number of bits given to both is reduced, the error can be absorbed by assigning appropriate bits to the spectrum correction vector. Also, from this, flexible bit allocation can be performed for each parameter.

(Embodiment 3) Next, a third embodiment of the present invention will be described with reference to the drawings. First, an encoding device will be described with reference to FIG. The same elements as those in the encoding device of the first embodiment shown in FIG. 1 are denoted by the same reference numerals, and duplicate description will be omitted. In FIG. 4, reference numeral 117 denotes a quantization level calculation unit which receives the pitch frequency and voiced / unvoiced judgment as inputs, and outputs the number of LPC coefficient quantization levels, the number of spectrum correction vector quantization levels, and the allocation of quantization bits. Reference numeral 118 denotes a variable level quantizer that receives an LPC coefficient and the number of LPC coefficient quantization levels and outputs a quantized LPC coefficient. Reference numeral 119 denotes a variable level quantizer which receives the spectrum correction vector and the number of quantization levels of the spectrum correction vector and outputs the quantization spectrum correction vector. Reference numeral 120 denotes a multiplexer that receives a pitch frequency, a voiced / unvoiced determination, a quantized LPC coefficient, a quantized frame power, and a quantized spectrum correction vector, and outputs a code. The residual information calculation unit 113B in this embodiment has a variable level quantizer 119 in place of the quantizer 112 shown in FIG.

Next, the operation will be described with reference to FIG.
The quantization level calculation unit 117 obtains the number of bits to be allocated to the LPC coefficient and the spectrum correction vector for each frame based on the pitch frequency and the voiced / unvoiced determination, and calculates the corresponding quantization level. For example, when the pitch frequency is high, the value of the number of harmonics L becomes small.
More bits can be allocated to the encoding of the C coefficient. On the contrary, there is no formant such as a silent part,
If the spectrum envelope changes gradually, it may be better to assign a larger number of bits to the spectrum correction vector. As described above, the total number of quantization bits assigned to each parameter, that is, the total number of quantization levels is fixed, and the LPC
By changing the allocation of the number of bits according to the information amount of the coefficient and the spectrum correction vector, more effective encoding can be performed. The variable level quantizer 118 quantizes LPC coefficients with the number of levels based on the number of allocated bits. Therefore, for example, if a plurality of vector quantizers having different codebook sizes are prepared, quantization can be performed efficiently. Similarly, the variable level quantizer 119 quantizes the spectrum correction vector with the number of levels based on the number of allocated bits. The multiplexer 120 encodes the quantized LPC coefficient and the quantized spectrum correction vector based on each bit allocation together with the pitch frequency, voiced / unvoiced decision, and quantized frame power.

Next, the decoding apparatus will be described with reference to FIG. The same elements as those of the decoding apparatus of the first embodiment shown in FIG. 2 are denoted by the same reference numerals, and the duplicate description will be omitted. In FIG. 5, reference numeral 210 denotes a quantization level number calculation unit which receives the pitch frequency and voiced / unvoiced determination as inputs, and outputs the number of LPC coefficient quantization levels, the number of spectrum correction vector quantization levels, and bit allocation. A demultiplexer 211 receives a code and a bit allocation, and outputs a pitch frequency, a voiced / unvoiced determination, a quantized LPC coefficient, a quantized frame power, and a quantized spectrum correction vector. Reference numeral 212 denotes a level variable inverse quantizer which receives a quantized LPC coefficient and the number of LPC coefficient quantization levels and outputs an inversely quantized LPC coefficient. A variable level inverse quantizer 213 receives the quantized spectrum correction vector and the number of quantization levels of the spectrum correction vector, and outputs the inversely quantized spectrum correction vector.

Next, the operation will be described with reference to FIG.
The demultiplexer 211 first outputs a pitch frequency, a voiced / unvoiced determination, and a quantized frame power from the code. In the quantization level number calculation unit 210, the number of LPC coefficient quantization levels, the number of spectrum correction vector quantization levels, and the number of allocated bits are set for each frame by means exactly equal to the quantization level number calculation unit 117 in the encoding apparatus of FIG. Is calculated. The demultiplexer 211 outputs a quantized LPC coefficient and a quantized spectrum correction vector based on the obtained number of allocated bits. In the level variable inverse quantizer 212, the LPC coefficient is quantized based on the number of quantization levels.
Dequantize the PC coefficients. This level variable inverse quantizer 2
12 is a variable level quantizer 118 in the encoding device.
Is an inverse quantizer corresponding to. Similarly, the variable level inverse quantizer 213 is an inverse quantizer corresponding to the variable level quantizer 119 in the encoding device, and inversely quantizes the spectrum correction vector based on the number of quantization levels of the spectrum correction vector. .

As described above, according to the present embodiment, the encoding apparatus includes a quantization level number calculator 117 and a corresponding variable level quantizer 118 for quantizing LPC coefficients.
, A variable level quantizer 119 for quantizing the spectrum correction vector, and a multiplexer 120, and the decoding apparatus is provided with a quantization level number calculation unit 11 of the encoding apparatus.
7; a demultiplexer 211 corresponding thereto; a variable level dequantizer 212 for dequantizing the quantized LPC coefficient; and a level for dequantizing the quantized spectrum correction vector. By providing the variable inverse quantizer 213, the allocation of the number of bits can be changed depending on the information amount of the LPC coefficient and the spectrum correction vector for each frame, more effective encoding can be performed, and the quality of decoded speech can be improved. be able to.

(Embodiment 4) Next, a fourth embodiment of the present invention will be described with reference to the drawings. First, an encoding device will be described with reference to FIG. The same elements as those in the encoding device of the first embodiment shown in FIG. 1 are denoted by the same reference numerals, and duplicate description will be omitted. In FIG. 6, 121 is L
A power vector calculation unit that receives a PC prediction residual as an input and outputs a power vector in a frame as an output. Reference numeral 122 denotes a quantizer that receives a power vector as an input and outputs a quantized power vector as an output. A multiplexer 123 receives a pitch frequency, a voiced / unvoiced determination, a quantized LPC coefficient, and a quantized power vector, and outputs a code. The residual information calculation unit 113C according to the present embodiment includes a power spectrum calculation unit 121 and a quantizer 122.

Next, the operation will be described with reference to FIG.
The power vector calculation unit 121 first divides the frame into a plurality of subframes in the LPC prediction residual,
The average power in each subframe is calculated, and the sum is output as a power vector of the frame. The quantizer 122 quantizes the power vector by an efficient quantization method such as vector quantization. Multiplexer 1
In 23, the pitch frequency, the voiced / unvoiced determination of each harmonic, the quantized LPC coefficient, and the quantized power vector obtained as information to be transmitted to the decoding device are efficiently encoded.

Next, the decoding apparatus will be described with reference to FIG. The same elements as those of the decoding apparatus of the first embodiment shown in FIG. 2 are denoted by the same reference numerals, and the duplicate description will be omitted. In FIG. 7, reference numeral 214 denotes a demultiplexer which receives a code and outputs a pitch frequency, a voiced / unvoiced determination, a quantized LPC coefficient, and a quantized power vector. 21
Reference numeral 5 denotes a quantizer which receives a quantized power vector as an input and outputs a dequantized power vector. 216 is M
A sound source generation unit that receives a BE mode and a power vector that has been inversely quantized, and outputs a sound source.

Next, the operation will be described with reference to FIG.
The demultiplexer 214 extracts a pitch frequency, a voiced / unvoiced determination, a quantized LPC coefficient, and a quantized power vector from the code transmitted from the encoder. The inverse quantizer 215 inversely quantizes the quantized power vector. The sound source generation unit 216 generates a sound source based on the power vector and the MBE combining mode. To generate a sound source, first, the frequency of each harmonic is obtained from the MBE synthesis mode, and the frequency of each harmonic is calculated. In addition, a voiced / unvoiced judgment is extracted, and a basic sound source signal is generated for each harmonic based on the extracted voiced / unvoiced judgment. Conventionally, a method is known in which a sine wave is used if the harmonic is voiced, and a white noise is used if the harmonic is unvoiced. Next, each fundamental signal is
An excitation signal is generated by dividing into the same number of subframes as the power vector calculation unit 121 of the encoding apparatus and multiplying the power of each subframe obtained from the power vector by the corresponding subframe of each harmonic. At this time,
To ensure that the sound source signal continues smoothly between subframes,
It is desirable to interpolate the power. In addition, MBE synthesis unit 2
When the synthesis is performed at step 09, since the modeling method differs between voiced and unvoiced for each harmonic, it is desirable to assume each model for voiced and unvoiced determination and remove the effect of the window function when calculating the spectrum. . As a method of removing the influence, a method of deconvolving the spectrum of the window function by assuming respective spectrum models for the amplitude of the sound source obtained from the frame power and the spectrum correction vector can be considered.

As described above, according to the present embodiment, the encoding apparatus includes the power vector calculating section 121, the quantizer 122 for quantizing the power vector calculating section 121, and the multiplexer 123, and the decoding apparatus includes the demultiplexer 214. , An inverse quantizer 215 for decoding a quantized power vector, and a sound source generation unit 2
By providing 16, even when the power greatly changes within a frame, such as at the time of voice rising,
The average power of each of the plurality of subframes in the frame is sent to the decoding device as a power vector, and the change in power can be reproduced at the time of generating the sound source, so that the quality of the decoded speech can be improved.

[0032]

As described above, according to the present invention, a coding apparatus is provided with a residual power spectrum calculating section, a residual harmonic power calculating section, a spectrum correction vector calculating section, and a method for quantizing a spectrum correction vector. By providing a quantizer for transforming, it is possible to calculate and encode a parameter representing a fine spectrum of each harmonic that could not be removed by the LPC analysis, and transmit the information to a decoding device as information.
Therefore, by providing the sound generator in consideration of the spectrum correction vector in the decoding device, the spectrum of the harmonic contained in the LPC prediction residual can be corrected at the time of decoding. Can be represented more accurately, and an excellent speech encoding / decoding apparatus capable of obtaining high decoded speech quality at the same bit rate as in the past can be realized.

Further, according to the present invention, the coding apparatus is provided with an inverse quantizer for inversely quantizing the quantized LPC coefficient and an inverse quantizer for inversely quantizing the quantized frame power. LPC inverse filter is applied by the obtained LPC coefficient,
By normalizing the LPC prediction residuals with the dequantized frame power, those quantization errors can be corrected by the spectral correction vector, and the number of bits given to the LPC coefficient and / or the frame power can be reduced. Also, the error can be absorbed by assigning appropriate bits to the spectrum correction vector. Also, from this, it is possible to flexibly assign bits to each parameter, and to realize an excellent speech encoding / decoding apparatus capable of obtaining high decoded speech quality at the same bit rate as in the related art.

Further, according to the present invention, the coding apparatus includes a quantization level number calculator, a corresponding level variable quantizer for quantizing LPC coefficients, and a level variable quantizer for quantizing a spectrum correction vector. A quantizer and a multiplexer are provided, and the decoding device is provided with the same quantization level number calculation unit as the quantization level number calculation unit of the encoding device, and the corresponding demultiplexer and quantized LPC coefficient are inverted. By providing a level-variable inverse quantizer for quantization and a level-variable inverse quantizer for inverse-quantizing the quantized spectrum correction vector, the number of bits can be allocated for each frame according to the information amount of the LPC coefficient and the spectrum correction vector. The superior speech encoding / decoding that can perform more effective encoding and achieve high decoded speech quality at the same bit rate as before The apparatus can be realized.

According to the present invention, a coding apparatus includes a power vector calculating section, a quantizer for quantizing the power vector calculating section,
By providing a multiplexer and providing a demultiplexer, an inverse quantizer for decoding a quantized power vector, and a sound source generation unit in a decoding device, the power greatly changes in a frame as at the time of rising of a voice. Even in such a case, the average power of each of a plurality of subframes in the frame is sent to the decoding device as a power vector, and the change in power can be reproduced at the time of generating the sound source, so that high decoded voice quality can be obtained at the same bit rate as before. An excellent speech encoding / decoding device obtained can be realized.

[Brief description of the drawings]

FIG. 1 is a block diagram of an encoding device according to a first embodiment of the present invention.

FIG. 2 is a block diagram of a decoding device according to a first embodiment and a second embodiment of the present invention;

FIG. 3 is a block diagram of an encoding device according to a second embodiment of the present invention.

FIG. 4 is a block diagram of an encoding device according to a third embodiment of the present invention.

FIG. 5 is a block diagram of a decoding device according to a third embodiment of the present invention.

FIG. 6 is a block diagram of an encoding device according to a fourth embodiment of the present invention.

FIG. 7 is a block diagram of a decoding device according to a fourth embodiment of the present invention.

FIG. 8 is a block diagram of a conventional encoding device.

FIG. 9 is a block diagram of a conventional decoding device.

[Explanation of symbols]

Reference Signs List 101 Pitch frequency estimation unit 102 Voiced / unvoiced judgment unit 103 LPC analysis unit 104 Quantizer 105 LPC inverse filter 106 Frame power calculation unit 107 Quantizer 108 Prediction residual normalization unit 109 Residual power spectrum calculation unit 110 Residual harmonic Power calculator 111 Spectrum correction vector calculator 112 Quantizer 113, 113A, 113B, 113C Residual information calculator 114 Multiplexer 115 Dequantizer 116 Dequantizer 117 Quantization level calculator 118 Level variable quantizer 119 Level variable quantizer 120 multiplexer 121 power vector calculator 122 quantizer 123 multiplexer 201 demultiplexer 202 mode determiner 203 inverse quantizer 204 spectrum enhancer 205 harmonic amplitude multiplier calculator 2 6 Dequantizer 207 Dequantizer 208 Sound source generator 209 MBE synthesizer 210 Quantization level number calculator 211 Demultiplexer 212 Level variable dequantizer 213 Level variable dequantizer 214 Demultiplexer 215 Dequantizer 216 Sound source generator

Claims (1)

(57) [Claims] 1. An encoding apparatus comprising: a pitch frequency estimating unit;
A voiced / unvoiced determination unit for voiced / unvoiced determination of each harmonic, and LP
A C (linear prediction) analyzer, a quantizer for quantizing LPC coefficients, an LPC inverse filter, a residual information calculator for extracting residual information other than at least residual frame power from the LPC predicted residual, A multiplexer for encoding the frequency, voiced / unvoiced determination, quantized LPC coefficient and residual information, wherein the decoding device is configured to determine the pitch frequency, voiced / unvoiced determination, quantized LPC coefficient and residual information from the code. A demultiplexer that decodes the LPC coefficient, a mode determination unit that determines the MBE combining mode, an inverse quantizer that inversely quantizes the quantized LPC coefficient to extract the LPC coefficient, and an amplitude from the LPC coefficient to each harmonic. It includes a harmonic amplitude multiplier calculating unit for obtaining a multiplier, and a sound source generating unit for generating a sound source signal from the pitch frequency and the voiced unvoiced frequency and residual information, and a MBE synthesis unit, encoder Definitive the residual information calculation unit, a frame-Pas
Power calculator and quantizer for quantizing frame power
A prediction residual normalization unit that normalizes the LPC prediction residual;
Power spectrum to calculate power spectrum of prediction residual
And the power of harmonics included in the residual
From the residual harmonic power calculator and the residual harmonic power,
Spectral correction vector to calculate vector correction vector
Calculator and quantization for quantizing the spectrum correction vector
A sound generator in the decoding apparatus includes a residual information
Power and spectrum calculated by the report calculation unit
A sound source is generated using the correction vector, and the encoding device and the decoding device
LPC coefficient and spectrum complement based on voiced and unvoiced decision
Calculate bit assignment for positive vector for each frame
And the number of quantization levels
Quantizer and inverse quantization for quantizing and inverse quantizing LPC coefficients
And inverse quantization of the detector and spectral correction vector
Quantizer and inverse quantizer calculate the quantization level
With different quantization levels for each frame given by the source
Performs quantization and inverse quantization, and also provides multiplexers and
The multiplexer calculates the number of quantization levels by the quantization level calculation unit.
Quantized LPC coefficients and
And quantize spectral correction vector
A speech encoding / decoding device characterized by the above-mentioned.